Top Mathematicians Form an Independent AI Advisory Group, With OpenAI as Its First Client

Nine of the world’s most prominent mathematicians announced the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) on September 21, in a guest post on Terence Tao’s blog. The group is hosted at the Institute for Advanced Study in Princeton, and its origin story is the notable part: OpenAI approached some of its members about establishing an external advisory board, and the mathematicians declined the shape of that offer and formed an independent group instead, then invited OpenAI to become its first client anyway. The lineup reads like a seminar table at the top of the field: Edward Witten (IAS), Timothy Gowers (Collège de France and Cambridge), Martin Hairer (EPFL and Imperial College), Melanie Matchett Wood (Harvard), Ravi Vakil (Stanford), Ulrike Tillmann (Oxford), Camillo De Lellis (IAS), Nikhil Srivastava (Berkeley), and François Charles (ENS-PSL). Between them they hold a Fields Medal, a Fields-adjacent Breakthrough catalog, and two of the discipline’s most influential blog voices.

Why OpenAI Went Looking for Mathematicians

The mandate, per the announcement, is “to advise AI companies on their interactions with mathematical research and with the mathematical community, including the responsible presentation and release of mathematical results.” The immediate workload explains the timing: AGMAI is currently advising OpenAI on how to coordinate the release of what OpenAI says is a large batch of significant results produced by its internal model, with commenters on the post citing claims of more than 100 long-standing open problems resolved. This is the same pipeline that produced the Navier-Stokes blow-up announcement on September 8, a result still awaiting community verification, and the sequence that followed: an open-letter wave, and yesterday’s essay by Po-Shen Loh on the same blog arguing the field must publicly decide its relationship with AI. OpenAI crossed a community norm “when Navier-Stokes was in sight,” as the post’s commenters put it, and the lab appears to have concluded that the cheapest fix is to hand the community a formal channel.

The Design: Independence as a Feature

AGMAI’s structural choices are a rebuttal to the standard corporate advisory board. Members take no payment. Recommendations are published. The group states plainly that its members “do not have decision making power at any AI company,” which cuts both ways: it protects independence and admits the group cannot force anything. Hosting at the Institute for Advanced Study, rather than at a lab or a consultancy, gives the arrangement an institutional home that outlives any single engagement. This is the model the math community has been gesturing at since 25 Fields Medalists declared severe misalignment risks in AI-for-math in September: not a boycott, but organized, paid-nothing, on-the-record scrutiny of how AI companies interact with the discipline.

For AI companies the offer is subtle. An external board you appoint is a signal you control; an independent group you can only petition is a signal you cannot. OpenAI choosing the latter, if that is what happened here, suggests the reputational cost of verification failures now exceeds the convenience of curating one’s own overseers.

What to Watch

The first test is the batch itself: whether the OpenAI results AGMAI shepherds into release survive verification, and whether the group’s publication recommendations (ordering, claiming conventions, credit for AI assistance) become a template other labs adopt. The second is precedent: if AGMAI works for mathematics, expect equivalents in physics, biology, and law, and watch which labs accept external versions versus building internal boards they control. The third is teeth: the group has no decision power and takes no money, which makes it influential exactly as long as labs find its seal valuable. The moment ignoring AGMAI gets cheaper than satisfying it, we will learn what the arrangement was really worth. The group is soliciting input from the broader community through a form on agmai.org, so the shape of that answer is partly up to the people who will use it.